A Comparison of Robot Navigation Algorithms

for an Unknown Goal

Russell Bayuk
Steven Ratering (faculty mentor)
Computer Science Department
University of Wisconsin – Eau Claire
Eau Claire, WI 54702
{bayukrj, raterisj}@uwec.edu

Abstract

Some robots have stationary bases and movable arms and grippers, while others
are mobile. Some robots’ movements are precisely programmed while others are more
autonomous. The robots we are interested in are both mobile and autonomous. A
common task for these robots is to go from one known location to another location whose
coordinates are given. The hard part is avoiding obstacles on the way. Another task is
finding a goal location when its coordinates are unknown. The robot senses it is at the
goal only after it arrives there. This is the problem we have investigated. We have
implemented several different algorithms and analyzed them in terms of how long the
robot takes to get to the goal and return to its starting location.
Introduction
A primary problem of robotics is getting a robot to move from one position to
another without bumping into any obstacles. This can be seen as either a path planning
problem or a navigation problem. The path planning problem[1,2,3] is solved by finding an
obstacle-free path through space and time from start configuration to goal configuration.
The obstacles may be stationary or moving. With sufficient a priori information this
problem can be solved before the robot starts to move, in fact one need not have a robot
to solve this problem. On the other hand, in the robot navigation problem, the robot
gathers information about the environment while it is moving. A fair amount of work has
been done with the robot navigation problem when the goal is known a priori.[4,5,6,7] Not
much has been done when the goal is not known a priori. The problem we address is
robot navigation in an environment with unknown stationary obstacles where the goal is
also unknown. After the robot reaches the goal it returns to its home cell. The robot
senses obstacles when it bumps into them and it senses the goal only after it arrives there.
The purpose is to determine the most efficient algorithm for navigating these unknown
environments.

The unknown environment is a rectangular room. This is to simulate the normal

environment of a robot, which is generally indoors. The room consists of a number of
cells. Between adjacent cells, there is the possibility of a wall. When a room is created a
cell is picked at random, that cell is assigned to be the goal cell. Since the room is
created with random wall placements, it is possible for parts of the room to be sealed off
by walls. The use of cells and walls can approximate any shape in a two-dimensional
environment. A closer approximation can be achieved by making the cells sufficiently
small.

The robot starts in the lower left hand corner facing to the “east” or to the right.
Its beginning coordinates are (0, 0). The only information that the robot knows about the
environment are the boundaries. It knows that it is contained in a rectangular area. It
knows how big the rectangle is and that it cannot leave the limits of the rectangle. It
knows that there is a goal inside the rectangle and that it is possible to reach the goal.
Even though it is possible for parts of the room to be inaccessible, the goal will not be in
these parts. The robot will know where the goal is only after it has reached the goal. It
cannot see the goal no matter how close it is until it is directly on top of it. Although the
robot doesn’t start out with much information about the environment, that doesn’t stop it
from learning as it goes along.

Learning is a critical step during the search of the room. Every cell the robot
visits, it remembers. Every wall that the robot bumps into gets recorded in a log. When
the robot passes from one cell to another without bumping into a wall, this is also
recorded. The robot then uses this knowledge to narrow down the places where the goal
may be. The path that the robot takes to uncover information about each cell is
determined by the combination of algorithms it is using.

1
Algorithms
There are many different types of navigation algorithms that a robot may use to
traverse space. Often these algorithms have similar processes but differ only slightly in
different situations. For example, when a robot hits a dead end, where does it go next?
Two different algorithms could have gotten the robot there using the exact same path, but
differ by how they choose the next step. Another way two algorithms could differ is in
how they choose the path back to the original start state after it has found the goal.

We found there to be three specific situations where there is a major difference

between navigation algorithms when choosing a path. Because of these situations, we
chose to split the algorithms into three different steps. The first step deals with the
generic situation, how to begin and continue the traversal of unknown space. This step
chooses which one of the neighboring cells to visit. The second situation is used when
the first situation cannot pick an unvisited cell to visit. This occurs when either, all the
adjacent unvisited cells are blocked by walls, or if all the adjacent cells have already been
visited. At this point, a decision needs to be made by the robot. It needs to choose a path
to continue its search of the unknown environment. The third step is when the robot has
found the goal and needs to proceed back to the initial starting point. There are several
different ways that the robot may choose to return to the origin.

There are several algorithms implemented for each of the steps. For the first step
there are two different algorithms. One is called “Follow the Right Wall.” This consists
of the robot following a wall similarly to a person walking along a path with their hand
always touching the wall on that side. A wall can be substituted with a cell that has
already been visited. The second algorithm implemented for the first step is called
“Right Left Sweep.” This algorithm is similar to “Follow the Right Wall” except when
the robot enters a cell that is adjacent to a boundary wall, the direction switches. For
example, if the robot is currently following the right wall and enters a cell with a
boundary wall the robot would start following the left wall. This would continue until it
enters another cell that is bordered by a boundary wall.

The implemented algorithms for the second step are as follows. The first
implementation is called “Find Closest Unvisited Cell.” This algorithm will consider the
cells the robot has not yet been in, and pick the one that is closest to the robot. The
second algorithm is “Find Closest Unvisited Cell to Start.” This is similar to the first
algorithm except that the base reference point is the start instead of the robot’s current
position.

The third algorithmic step, returning to the start after finding the goal, also has
multiple implementations. The first implementation is “Assume No Walls.” This is an
optimistic path finding algorithm that will plot a path back to the start assuming that all
unknown potential walls will not have walls. If a wall is found along the return path, a
new path needs to be calculated. The second is “Assume Walls.” This is a pessimistic
algorithm that will plot a path back to the start using only the information gathered. All
the unknowns are assumed to be a wall when traveling back to the start with this

2
algorithm. A new path will never need to be calculated because the return path is based
upon a path the robot knows exists. The third is “Assume Walls Unless Cell is Visited.”
This is in between the previous two algorithms. It will only assume an unknown is not a
wall if the cell it is attempting to enter has already been visited. Otherwise it will assume
that all unknowns are walls.

Twelve different algorithms can be created with the different combinations of the
above steps. We number the algorithms for later reference.

Algorithm Left Right Follow Closest to Closest to Assume No Assume No Walls if

Number Sweep Right Wall Robot Start Walls Walls Visited
1 X X X
2 X X X
3 X X X
4 X X X
5 X X X
6 X X X
7 X X X
8 X X X
9 X X X
10 X X X
11 X X X
12 X X X

We decided to run many different test cases pitting each of the algorithms against
each other to determine which algorithm, if any, is the best for a situation. The test case
consisted of the average of five hundred runs, for each algorithm, in random
environments with constant variables for its creation. The environment’s size was 10 by
10 cells. The initial percentage of obstacles was zero and we incremented it by 10 after
each test case, to maximum of 30 percent.

The results from these tests were split into three different categories. The robot
logged the number of bumps, turns and moves it took during the duration of each test.
With these different statistics we can apply weights to each type of movement and
anticipate the amount of time it would take different robots to reach the goal in the same
situation. The weights are multipliers applied to the different movement types. We used
three different sets of weights against our results, each adding up to a total of 15. The
first weight applied was five on each of the three movement types. This was to model the
robot that takes the same amount of time to do all three movements. The second is three
on one and six on the other two. This set demonstrates the robots that are quick at one of
the movements and slower at the other two. The third is nine on one and three on the
other two. This weight set describes the robots that are really slow at one of the
movements and fast at the other two movements.

After putting all this together, we are testing 12 different algorithms in four
different situations. Each of these situations yield seven different result sets, after being
weighted for the assortment of robots that are being represented. The analysis of these
different result sets should provide us with an accurate picture of what combinations

3
supply better results than others. These good algorithms will be further investigated later
to determine better combinations for robot navigation.

Results
The first test run was ten by ten cells with zero obstacles. The results that this test
presented were very interesting. It seems that the algorithms containing the pessimistic
end algorithm “Assume Walls” has higher movement counts and higher turn counts,
which almost double the movement and turn counts of the algorithms that don’t use
“Assume Walls.” Some characteristics of the algorithms are clearly showing in this
result set. Algorithms with “Follow the Right Wall” have a higher movement count than
the ones that have “Left Right Sweep.” Algorithm numbers 2 and 4 have the lowest total
count in all result sets except the two that put more weight on the move count.
Algorithms 5 and 7 took the longest in every single result set. Algorithms 1-4 are the
lowest and algorithms 9-12 are close behind and algorithms 5-8 take significantly longer.

The second test run increased the percentage of obstacles to ten percent. This
result set has additional information that wasn’t seen in the first test run. Since there are
obstacles in this test run we can find more information about the algorithms. Algorithm 1
had the most bumps while algorithm 7 had the least. This information may not be very
useful because the difference between the best and worst count was 2 bumps. Once again
the longest algorithms are 5-8; the “Assume Walls” algorithm appears to be responsible
for the slow completion time again. An interesting variance from the first set is that the

4
fastest four algorithms were the ones containing “Find Closest Unvisited Cell.” One odd
relation is the two slowest also contained “Find Closest Unvisited Cell” as part of their
algorithm.

The third test run raised the percentage of obstacles up to twenty percent. With
the increase in density of the obstacles some interesting patterns formed. The quickest
algorithms were once again the ones containing “Find Closest Unvisited Cell.” A
different pattern emerged in this run. “Left Right Sweep” combined with “Find Closest

5
Unvisited Cell to Start” took the longest. Also the weighting didn’t make much of a
difference in the result set. They may need to have more drastic differences in weights to
have an impact when the density of objects is higher. Algorithm 7 continued to rank
among the lowest for bumping, but it again had the highest count to finish.

The final run had the density at thirty percent of the cells. Most of the algorithms
took a relatively long time to complete this run. The only two algorithms that completed
this run comparatively fast were 2 and 10; they both contain the algorithms “Find closest
Unvisited Cell” and “Follow Right Wall”. Algorithm 7 once again took the longest to
complete the run, but this time it was joined by 3, 4, 5, 8, 11, and 12. Most of these
algorithms contain “Find Closest Unvisited Cell to Start”.

Conclusion
Many interesting patterns emerged in the result sets as the percentage of the
obstacles rose. Any algorithms containing “Find Closest Unvisited Cell to Start” or “Left
Right Sweep” became increasingly inefficient as the obstacle number rose. The “Assume
Walls” algorithm was very slow, but seemed to begin to hold its own as the number of
walls rose. It may become useful if the obstacle percentage rose even higher. The
winning algorithm is number 2. The pieces didn’t get hindered as much as the other
algorithms as the obstacle percentage rose, and it was consistently among the quickest
algorithms.

6
References
1
J. F.Canny. The Complexity of Robot Motion Planning. MIT Press, Cambridge, Massachusetts, 1988.
2
C. I. Connolly, J. B. Burns, and R. Weiss. Palth planning using Laplace’s Equations. In IEEE
International Conference on Robotics and Automation, pages 2102-2106. IEEE, May 1990.
3
V. J. Lumelsky and A. A. Stepanov. Path-planning strategies for a point mobile automaton moving
amidst unknown obstacles of arbitrary shape. Algorithmica, 2(4):403-430, 1987.
4
R. A. Brooks. Solving the find-path problem by good representation of free space. IEEE Transactions on
Systems, Man, and Cybernetics, SMC-13(3):190-197, March/April 1983
5
E. Gat. Reliable Goal-Directed Reactive Control of Autonomous Mobile Robots. PhD thesis, Virginia
Polytechnic Institute and State University, 1991.
6
S. Ratering and M. Gini. Robot Navigation in a Known Environment with Unknown Moving Obstacles.
In IEEE International Conference on Robotics and Automation, pages 25-30. IEEE, 1993.
7
M. G. Slack and D. P. Miller. Path planning through time and space in dynamic domains. In Sixth
National Conference on AI, pages 1067-1070, Seattle, WA, July 1987. AAAI.